. Author manuscript; available in PMC: 2024 May 10.

Published in final edited form as: Adv Neural Inf Process Syst. 2023 Dec;36:42258–42270.

Table 2:

Ablation of key components in CORP, measured by character error rate (CER) on RNN-decoded outputs to make all ablations comparable to the "No Language Model". The 95% confidence interval was computed via bootstrap resampling across 10 random seeds.

Ablation	CER% _{[95% CI]}
CORP	4.80 _{[4.68, 4.98]}
No Data Augmentation	5.87 _{[5.77, 6.01]}
No Replay Buffer	6.42 _{[6.33, 6.52]}
No GRU Backbone Training	7.00 _{[6.82, 7.19]}
No Language Model	19.22 _{[17.29, 22.18]}